The question NorthPeak Manufacturing asks is simple: which of its 40 machines will fail next week? For sixty years, industrial diagnostics answered that kind of question with rules. An expert writes a rule, a computer applies it. This lesson explains why data replaced rules, and what that change looks like on the daily readings of the NorthPeak fleet, the file data/clean/readings.csv presented on the previous page.
Each generation kept the previous one and added something.
| Period | Name | How the program decides | Example |
|---|---|---|---|
| 1960s to 1980s | Rule-based systems | A human writes every rule | "If temperature > 80, raise an alarm" |
| 1990s to 2000s | Machine learning | The program learns rules from examples | Spam filters, credit scoring |
| 2012 onward | Deep learning | Many layers learn from huge data | Image recognition, speech to text |
| 2017 onward | Generative AI | A model writes text, code or images | ChatGPT, Llama, Copilot |
A rule-based system is a long list of if-then statements. Experts write the rules. The computer only applies them. This works when the rules are few and clear.
Machine learning turns the idea around. You give the program many examples with the right answer. The program finds the rule by itself. A rule found this way is called a model.
Deep learning is machine learning with very large models. These models have many layers of simple units. They need a lot of data and strong computers. In 2012 a deep model won an image contest by a wide margin. That year changed the field.
Generative AI is deep learning that produces new content. A large language model, or LLM, writes the next word again and again. You will use one on your laptop this week.
data/clean/readings.csv has 14,600 rows: 40 machines times 365 days of 2025. Each row holds the load, temperature, vibration, pressure and power of one machine on one day, plus the column fault_next_7d, which says whether a fault followed within seven days.
A rule-based approach would say: "If temperature_c is above 80, a fault is coming." We tested this rule on the real file. It flags 133 rows. Only 29 of them are followed by a fault. And there are 1,085 faults in total. The rule misses 1,056 of them.
A machine learning approach reads all 14,600 rows. It looks at load, temperature, vibration, pressure and power together. It learns which mix of values comes before a fault. You will build this model in Weeks 5 and 6.
Beginners think machine learning replaced rules because rules are old. That is not the reason. Rules failed because the world has too many cases. Nobody can write them all. Data replaced rules because data already contains the cases. But data can be wrong or incomplete. Then the model learns the wrong thing. Good data comes first. That is why Weeks 2 to 4 are about data, not models.